NeuroBrix 0.6.15 on PyPI (
neurobrix), with nbxtensor 0.1.7 and nbxgrammar 0.1.1. Source: github.com/NeuroBrix/neurobrix, Apache License 2.0.
The NeuroBrix engine is the inference engine and its command line. A model ships as one .nbx container: its graph, its weights and everything needed to run it. The engine reads that container and runs it, whatever the model is (a chat model, a speech recogniser, a voice, an image or a video generator), with no per-model code.
How the products fit together, and what each one owns: Architecture.
| Engine | Flag | What it needs |
|---|---|---|
| Triton engine | --triton |
nbxtensor and Triton; no PyTorch. Installed by pip install neurobrix |
| PyTorch engine | --compiled |
PyTorch, installed for your hardware, then pip install "neurobrix[pytorch]" |
With no mode flag, a run takes the PyTorch engine where PyTorch is installed and the Triton engine where it is not, and its launch line says which. Prism plans memory and placement across the GPUs the machine has. Validated on NVIDIA and Apple GPUs, on Linux, Windows and macOS. The boundary: generic tensor work (kernels, device memory, allocator) lives in nbxtensor; the engine keeps the policy (Prism, the model families, the CLI, serving).
pip install neurobrix
neurobrix doctor
On a Mac with Apple Silicon, add the Metal build of Triton once with python -m nbxtensor install-metal. Every case: Install by hardware.
neurobrix hub --category LLM
neurobrix import TinyLlama/TinyLlama-1.1B-Chat-v1.0
neurobrix run --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --prompt "Write a Python function that reverses a string." --max-tokens 256
A model has one name, its hub name org/Model, for import, run, serve, upscale, inspect and remove. The models the hub serves are listed at neurobrix.org/models and by neurobrix hub.
neurobrix serve --model TinyLlama/TinyLlama-1.1B-Chat-v1.0
neurobrix chat
neurobrix stop
neurobrix serve keeps a model loaded in a background daemon; chat and run use it. With --http it also speaks the OpenAI API, with tools and constrained decoding through nbxgrammar. See the Serving guide and Serving and chat.
| Page | What it covers |
|---|---|
| Quickstart | From install to a first result, then a served model |
| Installation | Requirements, what lands on your disk |
| Install by hardware | NVIDIA, Apple Silicon, CPU, other GPUs |
| Your first hour | Step by step, one command and its expected result each |
| Run your first model | One command per kind of model |
| Inference and Serving | Running every family; keeping a model warm |
| The .nbx format, Runtime architecture, Zero principles | How the engine is built |
| CLI reference, Configuration, Troubleshooting | Every command, flag, variable and error message |
| Hardware support, Hardware profiles | What executes where, and how a machine is described |
| Benchmarks | Measured speed, the limits, and the method |
| Changelog | Every engine version, newest first |
© 2025–2026 Hocine Benkelaya. NeuroBrix engine released under the Apache License 2.0.
Developed by Hocine Benkelaya · Supported and funded by Neural Network Holding Ltd · Commercial distribution by WizWorks OÜ